closedhumanPR #42Negotiation

The Dealmakers: multi-attribute bargaining, Pareto frontier, agents learning to trade off price vs quantity

Multi-attribute negotiation scenario where buyer-seller pairs bargain simultaneously over price × quantity. Unlike price-only strategies, agents must explore both axes to reach the Pareto frontier — a joint optimum invisible to…

Author

cloud-post-code avatar

@cloud-post-code

github profile →
Status
Closed without merging
Opened on
Jun 27
Branch
hackathon/The_Dealmakers

Description

The pitch.

## Summary

Multi-attribute negotiation scenario where buyer-seller pairs bargain simultaneously over **price × quantity**. Unlike price-only strategies, agents must explore both axes to reach the Pareto frontier — a joint optimum invisible to `alternating_offers`.

**The integrative tension:** buyers have a convex sweet-spot on quantity (they dislike over-ordering as much as under-ordering), sellers prefer larger volumes. Because their ideal quantities differ and both utility functions respond in opposite directions, a Pareto-optimal contract exists that is strictly better for *both* parties than either opening position.

## What's in the PR

- **`packages/nest-core/nest_core/scenarios_builtin/dealmakers.py`** — `DealParams` (deterministic utility params per role), `DealmakerAgent` (buyer/seller with cross-session EMA learning), and `dealmakers_factory`
- **`scenarios/dealmakers.yaml`** — 8 buyer-seller pairs, seed 13, 10 rounds
- **`packages/nest-core/nest_core/scenarios.py`** — registers `dealmakers`
- **`packages/nest-core/nest_core/validators.py`** — 3 validators: `sessions_closed`, `pareto_progress` (quantity axis must move), `learning_emitted` (every closer emits `learn:` line)
- **`packages/nest-core/tests/test_dealmakers.py`** — 19 tests covering params, strategies, learning, integration, determinism, validators, and Pareto exploration

## Three built-in strategies

| Strategy | Mechanism |
|---|---|
| `logrolling` | Stay on iso-utility contour, maximise estimated opponent utility |
| `ttt_mirror` | Mirror the opponent's last per-axis movement |
| `nash_split` | Propose midpoint biased toward own reservation with rounds |

## Cross-session learning

Each agent maintains a per-partner EMA estimate of opponent weight on price vs quantity, updated after every closed session. The agent emits a `learn:` trace line recording the updated estimate. In subsequent sessions the estimate biases counter-proposals toward the opponent's revealed preferences.

## Validators

1. **`dealmakers_sessions_closed`** — every pair that exchanged offers must have exactly one `close:` line
2. **`dealmakers_pareto_progress`** — at least one pair must move the quantity axis (proving Pareto exploration beyond price-only)
3. **`dealmakers_learning_emitted`** — every agent that closed a session must emit a `learn:` line

## Test plan

- [x] `uv sync` — clean
- [x] `uv run ruff check .` — 0 errors
- [x] `uv run ruff format --check .` — 0 diffs
- [x] `uv run pyright` — 0 new errors (1 pre-existing in `test_pareto_strategies.py`)
- [x] `uv run pytest -v` — 643 passed, 1 skipped

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Try it

Open PR on GitHubView diff

Checkout locally

git fetch origin pull/42/head:pr-42
git checkout pr-42